WO2008155124A2

WO2008155124A2 - Load balancing

Info

Publication number: WO2008155124A2
Application number: PCT/EP2008/004963
Authority: WO
Inventors: Vladimir Grouzdev
Original assignee: Virtuallogix Sa
Priority date: 2007-06-19
Filing date: 2008-06-19
Publication date: 2008-12-24
Also published as: EP2527980A2; EP2006770A1; EP2527980A3; US20080320489A1; US8341630B2; EP2006770B1; WO2008155124A3

Abstract

In a preferred embodiment, the present invention provides a method of load balancing in a data processing system comprising a plurality of physical CPUs and a plurality of virtual CPUs, the method comprising: mapping one or more virtual CPUs to each of said physical CPUs; and dynamically adapting the mapping depending on the load of said physical CPUs.

Description

LOAD BALANCING

Background

A virtual machine is a self-contained execution environment that behaves as if it is a separate computer and which can run its own operating system. Virtual machines provide Virtual CPUs (VCPUs) to clients or "guests", and each VCPU runs on a dedicated physical CPU. A VCPU is a representation of a physical processor within a Virtual Machine. In conventional systems, the mapping between virtual and physical CPUs is static.

If multiple CPUs are available to a client or "guest", the guest tasks are spread between the CPUs. This is preferably done such that the available resources are used in the most efficient way and computing time is decreased. This process is generally referred to as "load balancing".

Conventional load balancing algorithms may be insufficient. Let us consider, for example, the sharing of a plurality of physical CPUs between dedicated real-time software and generic server software. Let us assume an UP (uniprocessor) execution environment (e.g. LINUX) running the real-time software on CPU 0, and an SMP (symmetric multiprocessing) execution environment (LINUX) running a server software on CPUs 0-3. In this example, CPU 0 is shared between the the real-time software and the server software. The dedicated real-time has a higher scheduling priority. In this example, the SMP load balancer does not take into account the realtime activity on CPU 0. This may skew the SMP load balancing. The present invention aims to address this and other problems of conventional load balancing. In particular, but not exclusively, the present invention is concerned with better balancing the load of physical CPUs in a computer system comprising physical and virtual CPUs.

Summary of the Invention

The invention is recited by the independent claims. Preferred features are recited by the dependent claims.

Brief Description of the Drawings

Fig. 1 illustrates schematically the architecture of a system to which the present invention can be applied; and

Figs. 2 to 5 illustrate screen shots showing CPU usage statistics with and without load balancing in accordance with an embodiment of the invention.

Description of Exemplary Embodiments of the Invention

Introduction

Fig. 1 illustrates schematically the architecture of a system comprising physical

CPUs, virtual CPUs, and client (guest) applications, to which the present invention can be applied.

The present invention is based on the realisation that the overall performance of a system as shown in Fig. 1 can be improved by automatically balancing the load on physical CPUs attributed to the same SMP guest. In particular, the idea is to balance the physical CPUs load by migrating a virtual CPU (VCPU) from one physical CPU to another. Although this affects a statical VCPU to CPU assignment for a UP guest, the VCPUs to CPUs mapping within a fixed set of CPUs is transparent for SMP guests such that a scheduler which implements the present invention can decide which VCPU is to run on which physical CPU within the fixed CPU set. In other words, if an SMP guest is running on two physical CPUs (a & b), the scheduler creates two VCPUs (x & y) and maps them to the physical CPUs. In this example, two equivalent mappings are possible:

(1) VCPU_x → CPU₃ and VCPU_y → CPUb

(2) VCPU_x → CPU_b and VCPU_y → CPU_a

The scheduler is able to dynamically choose one of these mappings depending on CPU loads. The mapping switch results in a "swapping" of VCPUs, i.e. in two VCPUs migrating from one physical CPU to another. Such an operation is fully transparent for the guest and does not change a fixed physical set of CPUs assigned to the guest.

By implementing such a load balancing mechanism, it is possible to at least partially resolve the above described SMP load balancing skewing problem by migrating a server VCPU running on CPUo, f°^r example, to another CPU when the real-time activity is high and this VCPU is loaded. In order to improve the overall performance by such a migration, an underloaded CPU_n (n > 0, in this example)

must be found and VCPU running on CPU_n must migrate to CPUQ. This solution is partial only in that it does not work when the system is heavily loaded, i.e. when all physical CPUs are fully loaded. However, as such a situation is rare in practice, it is acceptable.

Migration criteria

Regularly, at a given time period, the scheduler calculates the load of each physical

CPU (LOAD_n, n=0..N) and all VCPUs running on each physical CPU (VLOAD^m, m=0..M). More particularly, two loads are calculated for each VCPU: a "positive" load and a "negative" load.

The positive load for a given VCPU_{n m} is equal to the actual VCPU load:

VPLOAD_n)m = VLOAD_n,_m

The negative load for a given VCPU_{n m} is equal to a sum of the loads of all other VCPUs running on the same physical CPU:

VNLOAD_n>m = ∑VLOAD_n)i i=0..M, i ≠ m

A physical CPU is considered overloaded if its load is above a predetermined threshold:

LOAD_n > LOAD_over A physical CPU is considered underloaded if its load is below a predetermined threshold:

LOAD_n < LOAD_under

Load balancing is only applied to a pair of CPUs in which one CPU is overloaded and other CPU is underloaded:

CPUi _^ CPU_j where CPUi ≥ LOAD_over and CPU_j < LOAD_und_er

The load balancing comprises finding two unbalanced VCPUs of the same SMP guest running on CPUj and CPU_j such that:

VPLOAD_i)k > VPLOADμ

and swapping these VCPUs across physical CPUs.

Because a VCPU migration, in terms of processing power, is a quite expensive operation, the migration criteria is adjusted by introducing a positive migration threshold: VPLOAD_i;k - VPLO ADy > MIGR _P OS JVTMARK

In addition, the migration criteria takes into account the negative load of the overloaded emigrant:

VNL0AD_i>k > MlGRJNEG JVTMARK

The negative load water mark avoids unnecessary migrations when the CPU overloading is not caused by a simultaneous activity of multiple guests, but rather by a single guest monopolizing the physical CPU.

Specific Implementation

A mymips program has been used to demonstrate skewing in the SMP load balancing of a guest operating system. The mymips program permanently calculates the program execution speed (MIPS) and prints out the calculation results on console.

mymips provides the following result when running on a SMP Linux guest with a dedicated single physical CPU:

min/max/ave: 258/258/258

The results above and below were obtained on a DELL D820 Dual Core 1.8 MHz Laptop. Two mymips programs provide the following result when running on an SMP Linux with two dedicated CPUs:

min/max/ave: 257/259/258 min/max/ave: 258/259/258

A basic configuration which can be used to implement the load balancing mechanism in accordance with an embodiment of the invention comprises two SMP Linux guests sharing two physical CPUs, hi order to obtain an unbalanced load on such a configuration, guests have been running on a conventional system without load balancing mechanism. Two mymips programs running simultaneously on each guest provide the following results:

min/max/ave: 101/258/190 min/max/ave: 92/257/190

This shows about 25% of performance hit comparing to a single SMP Linux which performs a load balancing (at OS level) across multiple CPUs. The performance hit is due to sporadic mymips migrations from one CPU to another. Such a migration randomly runs mymips on the same processor.

This practical result is fully in line with a theoretical determination. Because of a random nature of migrations, the probability of running both mymips on the same CPU is 0.5. Thus, an expected performance hit is 0.25 because when running two programs on the same CPU, only a half of the CPU power is available. Figs. 2 illustrates a screen shot showing CPU usage statistics without load balancing in accordance with an embodiment of the invention. The screen shot represents the scenario described above.

When running the same load with load balancing enabled, the performance is close to a single SMP Linux.

min/max/ave: 254/257/256 min/max/ave: 250/256/255

The load balancing compensates sporadic migrations oϊmymips (from one VCPU to another) caused by the Linux SMP scheduler. In other words, the scheduler tries to execute heavy loaded VCPUs on different physical CPUs.

Figs. 3 illustrates a screen shot showing CPU usage statistics with load balancing in accordance with an embodiment of the invention. The screen shot represents the scenario described above.

In order to confirm the above theoretical conclusion, a Linux kernel compilation was used as a variable load. Two compilations were running in parallel on two Linux guests in the following three configurations:

(1) Two Linux guests each running on a dedicated CPU (2) Two Linux guests sharing two CPUs without load balancing (3) Two Linux guests sharing two CPUs with load balancing

Each time, the duration of compilation was measured. Results corresponding to different Linux kernel compilations are shown below.

(1.1) l lm21.046s

(1.2) 8m34.204s

(2.1) 16m4.272s

(2.2) 12m20.575s

(3.1) 13m51.974s

(3.2) 10m32.467s

The performance hit on the system without load balancing (2) is about 40%, while the performance hit on the system with load balancing (3) is about 20%. Accordingly, the load balancing improves a performance degradation caused by a transparent CPU sharing among multiple SMP guests.

Figs. 4 and 5 illustrate screen shots showing CPU usage statistics with and without load balancing in accordance with an embodiment of the invention. The screen shots represent the scenarios described above. The above results were obtained using the following load balancing parameters:

PERIOD = 10 milliseconds LOAD_under = 80%

LOAD_Over = 100%

MIGR_POS_WTMARK= 5% MIGR_NEG JVTMARK = 5%

It may be possible to achieve even better result on a variable load by modifying these parameters.

Other aspects and embodiments

It will be clear from the forgoing that the above-described embodiments are only examples, and that other embodiments are possible and included within the scope of the invention as determined from the claims.

Claims

1. A method of load balancing in a data processing system comprising a plurality of physical CPUs and a plurality of virtual CPUs, the method comprising: mapping one or more virtual CPUs to each of said physical CPUs; and dynamically adapting the mapping depending on the load of said physical CPUs.

2. The method of claim 1 , comprising: running a multiprocessor operation involving at least two physical CPUs; and assigning at least two virtual CPUs to said multiprocessor operation, for executing said multiprocessor operation; and mapping the at least two virtual CPUs to said at least two physical CPUs, respectively, wherein the mapping depends on the load of said at least two physical CPUs.

3. The method of claim 2, comprising: swapping the mapping of the two virtual CPUs to the at least two physical CPUs, respectively, in response to a change of load of at least one of the two physical CPUs.

4. The method of claim 2 or 3, comprising: determining the load of said two virtual CPUs; determining whether the load of said two physical CPUs is above a first threshold or below a second threshold; and swapping the mapping of the two virtual CPUs to the two physical CPUs, respectively, if the following conditions are met:

(i) the load of one of the two physical CPUs is above said first threshold, (ii) the load of the other one of the two physical CPUs is below said second threshold, and

(iii) the load of a virtual CPU mapped to the physical CPU whose load is above said first threshold is higher than the load of a virtual CPU mapped to the physical CPU whose load is below said second threshold.

5. The method of claim 4, comprising: swapping the mapping only if the difference of loads of said two virtual CPUs is above a predetermined threshold.

6. The method of claim 4 or 5, comprising: determining, for a given virtual CPU, a negative load indicative of the load of all other virtual CPUs that are allocated to the same physical CPU as the given virtual CPU; and swapping the mapping only if the negative load is above a predetermined threshold.

7. The method of claim 2, comprising: running an operation of a high scheduling priority on one of said at least two physical CPUs, wherein said multiprocessor operation has a relatively lower scheduling priority.

8. The method of claim 7, wherein the operation of a higher scheduling priority is executed by a dedicated real-time software.

9. The method of claim 7 or 8, wherein said multiprocessor operation is executed by a generic server software.

10. Computer system, arranged to perform the method of any preceding claim.

11. Computer program product, comprising machine-readable code which, when executed by a data processing device, executes the method of any of claims 1 to 9.